Human attention is limited, while many digital systems require users to manage multiple evolving processes at the same time. Although reinforcement learning has shown potential for automating attention management in simple dual-task settings, it remains unknown how these benefits change when the number of concurrent tasks increases. This thesis investigates the scalability of a reinforcement learning-based Attention Management System that is designed to manage forced task switching in a continuous environment where the task count ranges from two to nine. To ensure that the learned policy remains compatible with human limitations, the supervisor was trained using a cognitively constrained simulated user that included partial observability and post-switch motor inertia. In addition, the system applies a serialized pre-cue and forced-commit protocol, which enforces minimum dwell times to respect human reorientation needs and service-time limits.
A within-subject user study compared the performance and behavioral strategies of the automated system against a self-managed baseline condition. The evaluation identified two distinct performance regimes. In the effective regime where the task count was five or fewer, the system performed better than the manual baseline, as it reduced the rate of performance degradation by adopting a proactive strategy that increased switching pace and prioritized riskier tasks without increasing user reaction times. However, at higher task counts of six or more, the system exhibited a stability limit termed the "mute barrier," where the supervisor ceased scheduling switches, which led to failure due to task starvation. These findings provide empirical evidence that automated supervision helps address human multitasking capacity limits in moderate scaling scenarios while also highlighted specific stability challenges that should be addressed to support high-concurrency environments.
